2 * G.723.1 compatible decoder
3 * Copyright (c) 2006 Benjamin Larsson
4 * Copyright (c) 2010 Mohamed Naufal Basheer
6 * This file is part of FFmpeg.
8 * FFmpeg is free software; you can redistribute it and/or
9 * modify it under the terms of the GNU Lesser General Public
10 * License as published by the Free Software Foundation; either
11 * version 2.1 of the License, or (at your option) any later version.
13 * FFmpeg is distributed in the hope that it will be useful,
14 * but WITHOUT ANY WARRANTY; without even the implied warranty of
15 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
16 * Lesser General Public License for more details.
18 * You should have received a copy of the GNU Lesser General Public
19 * License along with FFmpeg; if not, write to the Free Software
20 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
25 * G.723.1 compatible decoder
28 #define BITSTREAM_READER_LE
29 #include "libavutil/audioconvert.h"
30 #include "libavutil/lzo.h"
31 #include "libavutil/opt.h"
35 #include "acelp_vectors.h"
36 #include "celp_filters.h"
37 #include "celp_math.h"
38 #include "g723_1_data.h"
40 typedef struct g723_1_context {
44 G723_1_Subframe subframe[4];
45 enum FrameType cur_frame_type;
46 enum FrameType past_frame_type;
48 uint8_t lsp_index[LSP_BANDS];
52 int16_t prev_lsp[LPC_ORDER];
53 int16_t prev_excitation[PITCH_MAX];
54 int16_t excitation[PITCH_MAX + FRAME_LEN + 4];
55 int16_t synth_mem[LPC_ORDER];
56 int16_t fir_mem[LPC_ORDER];
57 int iir_mem[LPC_ORDER];
65 int pf_gain; ///< formant postfilter
66 ///< gain scaling unit memory
68 int16_t audio[FRAME_LEN + LPC_ORDER];
69 int16_t prev_data[HALF_FRAME_LEN];
70 int16_t prev_weight_sig[PITCH_MAX];
73 int16_t hpf_fir_mem; ///< highpass filter fir
74 int hpf_iir_mem; ///< and iir memories
75 int16_t perf_fir_mem[LPC_ORDER]; ///< perceptual filter fir
76 int16_t perf_iir_mem[LPC_ORDER]; ///< and iir memories
78 int16_t harmonic_mem[PITCH_MAX];
81 static av_cold int g723_1_decode_init(AVCodecContext *avctx)
83 G723_1_Context *p = avctx->priv_data;
85 avctx->channel_layout = AV_CH_LAYOUT_MONO;
86 avctx->sample_fmt = AV_SAMPLE_FMT_S16;
90 avcodec_get_frame_defaults(&p->frame);
91 avctx->coded_frame = &p->frame;
93 memcpy(p->prev_lsp, dc_lsp, LPC_ORDER * sizeof(*p->prev_lsp));
99 * Unpack the frame into parameters.
101 * @param p the context
102 * @param buf pointer to the input buffer
103 * @param buf_size size of the input buffer
105 static int unpack_bitstream(G723_1_Context *p, const uint8_t *buf,
110 int temp, info_bits, i;
112 init_get_bits(&gb, buf, buf_size * 8);
114 /* Extract frame type and rate info */
115 info_bits = get_bits(&gb, 2);
117 if (info_bits == 3) {
118 p->cur_frame_type = UNTRANSMITTED_FRAME;
122 /* Extract 24 bit lsp indices, 8 bit for each band */
123 p->lsp_index[2] = get_bits(&gb, 8);
124 p->lsp_index[1] = get_bits(&gb, 8);
125 p->lsp_index[0] = get_bits(&gb, 8);
127 if (info_bits == 2) {
128 p->cur_frame_type = SID_FRAME;
129 p->subframe[0].amp_index = get_bits(&gb, 6);
133 /* Extract the info common to both rates */
134 p->cur_rate = info_bits ? RATE_5300 : RATE_6300;
135 p->cur_frame_type = ACTIVE_FRAME;
137 p->pitch_lag[0] = get_bits(&gb, 7);
138 if (p->pitch_lag[0] > 123) /* test if forbidden code */
140 p->pitch_lag[0] += PITCH_MIN;
141 p->subframe[1].ad_cb_lag = get_bits(&gb, 2);
143 p->pitch_lag[1] = get_bits(&gb, 7);
144 if (p->pitch_lag[1] > 123)
146 p->pitch_lag[1] += PITCH_MIN;
147 p->subframe[3].ad_cb_lag = get_bits(&gb, 2);
148 p->subframe[0].ad_cb_lag = 1;
149 p->subframe[2].ad_cb_lag = 1;
151 for (i = 0; i < SUBFRAMES; i++) {
152 /* Extract combined gain */
153 temp = get_bits(&gb, 12);
155 p->subframe[i].dirac_train = 0;
156 if (p->cur_rate == RATE_6300 && p->pitch_lag[i >> 1] < SUBFRAME_LEN - 2) {
157 p->subframe[i].dirac_train = temp >> 11;
161 p->subframe[i].ad_cb_gain = FASTDIV(temp, GAIN_LEVELS);
162 if (p->subframe[i].ad_cb_gain < ad_cb_len) {
163 p->subframe[i].amp_index = temp - p->subframe[i].ad_cb_gain *
170 p->subframe[0].grid_index = get_bits1(&gb);
171 p->subframe[1].grid_index = get_bits1(&gb);
172 p->subframe[2].grid_index = get_bits1(&gb);
173 p->subframe[3].grid_index = get_bits1(&gb);
175 if (p->cur_rate == RATE_6300) {
176 skip_bits1(&gb); /* skip reserved bit */
178 /* Compute pulse_pos index using the 13-bit combined position index */
179 temp = get_bits(&gb, 13);
180 p->subframe[0].pulse_pos = temp / 810;
182 temp -= p->subframe[0].pulse_pos * 810;
183 p->subframe[1].pulse_pos = FASTDIV(temp, 90);
185 temp -= p->subframe[1].pulse_pos * 90;
186 p->subframe[2].pulse_pos = FASTDIV(temp, 9);
187 p->subframe[3].pulse_pos = temp - p->subframe[2].pulse_pos * 9;
189 p->subframe[0].pulse_pos = (p->subframe[0].pulse_pos << 16) +
191 p->subframe[1].pulse_pos = (p->subframe[1].pulse_pos << 14) +
193 p->subframe[2].pulse_pos = (p->subframe[2].pulse_pos << 16) +
195 p->subframe[3].pulse_pos = (p->subframe[3].pulse_pos << 14) +
198 p->subframe[0].pulse_sign = get_bits(&gb, 6);
199 p->subframe[1].pulse_sign = get_bits(&gb, 5);
200 p->subframe[2].pulse_sign = get_bits(&gb, 6);
201 p->subframe[3].pulse_sign = get_bits(&gb, 5);
202 } else { /* 5300 bps */
203 p->subframe[0].pulse_pos = get_bits(&gb, 12);
204 p->subframe[1].pulse_pos = get_bits(&gb, 12);
205 p->subframe[2].pulse_pos = get_bits(&gb, 12);
206 p->subframe[3].pulse_pos = get_bits(&gb, 12);
208 p->subframe[0].pulse_sign = get_bits(&gb, 4);
209 p->subframe[1].pulse_sign = get_bits(&gb, 4);
210 p->subframe[2].pulse_sign = get_bits(&gb, 4);
211 p->subframe[3].pulse_sign = get_bits(&gb, 4);
218 * Bitexact implementation of sqrt(val/2).
220 static int16_t square_root(int val)
222 return (ff_sqrt(val << 1) >> 1) & (~1);
226 * Calculate the number of left-shifts required for normalizing the input.
228 * @param num input number
229 * @param width width of the input, 15 or 31 bits
231 static int normalize_bits(int num, int width)
240 i= width - av_log2(num) - 1;
246 #define normalize_bits_int16(num) normalize_bits(num, 15)
247 #define normalize_bits_int32(num) normalize_bits(num, 31)
248 #define dot_product(a,b,c,d) (ff_dot_product(a,b,c)<<(d))
251 * Scale vector contents based on the largest of their absolutes.
253 static int scale_vector(int16_t *vector, int length)
255 int bits, scale, max = 0;
258 const int16_t shift_table[16] = {
259 0x0001, 0x0002, 0x0004, 0x0008, 0x0010, 0x0020, 0x0040, 0x0080,
260 0x0100, 0x0200, 0x0400, 0x0800, 0x1000, 0x2000, 0x4000, 0x7fff
263 for (i = 0; i < length; i++)
264 max = FFMAX(max, FFABS(vector[i]));
266 max = FFMIN(max, 0x7FFF);
267 bits = normalize_bits(max, 15);
268 scale = shift_table[bits];
270 for (i = 0; i < length; i++) {
271 av_assert2(av_clipl_int32(vector[i] * (int64_t)scale << 1) == vector[i] * (int64_t)scale << 1);
272 vector[i] = (vector[i] * scale) >> 3;
279 * Perform inverse quantization of LSP frequencies.
281 * @param cur_lsp the current LSP vector
282 * @param prev_lsp the previous LSP vector
283 * @param lsp_index VQ indices
284 * @param bad_frame bad frame flag
286 static void inverse_quant(int16_t *cur_lsp, int16_t *prev_lsp,
287 uint8_t *lsp_index, int bad_frame)
290 int i, j, temp, stable;
292 /* Check for frame erasure */
299 lsp_index[0] = lsp_index[1] = lsp_index[2] = 0;
302 /* Get the VQ table entry corresponding to the transmitted index */
303 cur_lsp[0] = lsp_band0[lsp_index[0]][0];
304 cur_lsp[1] = lsp_band0[lsp_index[0]][1];
305 cur_lsp[2] = lsp_band0[lsp_index[0]][2];
306 cur_lsp[3] = lsp_band1[lsp_index[1]][0];
307 cur_lsp[4] = lsp_band1[lsp_index[1]][1];
308 cur_lsp[5] = lsp_band1[lsp_index[1]][2];
309 cur_lsp[6] = lsp_band2[lsp_index[2]][0];
310 cur_lsp[7] = lsp_band2[lsp_index[2]][1];
311 cur_lsp[8] = lsp_band2[lsp_index[2]][2];
312 cur_lsp[9] = lsp_band2[lsp_index[2]][3];
314 /* Add predicted vector & DC component to the previously quantized vector */
315 for (i = 0; i < LPC_ORDER; i++) {
316 temp = ((prev_lsp[i] - dc_lsp[i]) * pred + (1 << 14)) >> 15;
317 cur_lsp[i] += dc_lsp[i] + temp;
320 for (i = 0; i < LPC_ORDER; i++) {
321 cur_lsp[0] = FFMAX(cur_lsp[0], 0x180);
322 cur_lsp[LPC_ORDER - 1] = FFMIN(cur_lsp[LPC_ORDER - 1], 0x7e00);
324 /* Stability check */
325 for (j = 1; j < LPC_ORDER; j++) {
326 temp = min_dist + cur_lsp[j - 1] - cur_lsp[j];
329 cur_lsp[j - 1] -= temp;
334 for (j = 1; j < LPC_ORDER; j++) {
335 temp = cur_lsp[j - 1] + min_dist - cur_lsp[j] - 4;
345 memcpy(cur_lsp, prev_lsp, LPC_ORDER * sizeof(*cur_lsp));
349 * Bitexact implementation of 2ab scaled by 1/2^16.
351 * @param a 32 bit multiplicand
352 * @param b 16 bit multiplier
354 #define MULL2(a, b) \
358 * Convert LSP frequencies to LPC coefficients.
360 * @param lpc buffer for LPC coefficients
362 static void lsp2lpc(int16_t *lpc)
364 int f1[LPC_ORDER / 2 + 1];
365 int f2[LPC_ORDER / 2 + 1];
368 /* Calculate negative cosine */
369 for (j = 0; j < LPC_ORDER; j++) {
370 int index = lpc[j] >> 7;
371 int offset = lpc[j] & 0x7f;
372 int64_t temp1 = cos_tab[index] << 16;
373 int temp2 = (cos_tab[index + 1] - cos_tab[index]) *
374 ((offset << 8) + 0x80) << 1;
376 lpc[j] = -(av_clipl_int32(((temp1 + temp2) << 1) + (1 << 15)) >> 16);
380 * Compute sum and difference polynomial coefficients
381 * (bitexact alternative to lsp2poly() in lsp.c)
383 /* Initialize with values in Q28 */
385 f1[1] = (lpc[0] << 14) + (lpc[2] << 14);
386 f1[2] = lpc[0] * lpc[2] + (2 << 28);
389 f2[1] = (lpc[1] << 14) + (lpc[3] << 14);
390 f2[2] = lpc[1] * lpc[3] + (2 << 28);
393 * Calculate and scale the coefficients by 1/2 in
394 * each iteration for a final scaling factor of Q25
396 for (i = 2; i < LPC_ORDER / 2; i++) {
397 f1[i + 1] = f1[i - 1] + MULL2(f1[i], lpc[2 * i]);
398 f2[i + 1] = f2[i - 1] + MULL2(f2[i], lpc[2 * i + 1]);
400 for (j = i; j >= 2; j--) {
401 f1[j] = MULL2(f1[j - 1], lpc[2 * i]) +
402 (f1[j] >> 1) + (f1[j - 2] >> 1);
403 f2[j] = MULL2(f2[j - 1], lpc[2 * i + 1]) +
404 (f2[j] >> 1) + (f2[j - 2] >> 1);
409 f1[1] = ((lpc[2 * i] << 16 >> i) + f1[1]) >> 1;
410 f2[1] = ((lpc[2 * i + 1] << 16 >> i) + f2[1]) >> 1;
413 /* Convert polynomial coefficients to LPC coefficients */
414 for (i = 0; i < LPC_ORDER / 2; i++) {
415 int64_t ff1 = f1[i + 1] + f1[i];
416 int64_t ff2 = f2[i + 1] - f2[i];
418 lpc[i] = av_clipl_int32(((ff1 + ff2) << 3) + (1 << 15)) >> 16;
419 lpc[LPC_ORDER - i - 1] = av_clipl_int32(((ff1 - ff2) << 3) +
425 * Quantize LSP frequencies by interpolation and convert them to
426 * the corresponding LPC coefficients.
428 * @param lpc buffer for LPC coefficients
429 * @param cur_lsp the current LSP vector
430 * @param prev_lsp the previous LSP vector
432 static void lsp_interpolate(int16_t *lpc, int16_t *cur_lsp, int16_t *prev_lsp)
435 int16_t *lpc_ptr = lpc;
437 /* cur_lsp * 0.25 + prev_lsp * 0.75 */
438 ff_acelp_weighted_vector_sum(lpc, cur_lsp, prev_lsp,
439 4096, 12288, 1 << 13, 14, LPC_ORDER);
440 ff_acelp_weighted_vector_sum(lpc + LPC_ORDER, cur_lsp, prev_lsp,
441 8192, 8192, 1 << 13, 14, LPC_ORDER);
442 ff_acelp_weighted_vector_sum(lpc + 2 * LPC_ORDER, cur_lsp, prev_lsp,
443 12288, 4096, 1 << 13, 14, LPC_ORDER);
444 memcpy(lpc + 3 * LPC_ORDER, cur_lsp, LPC_ORDER * sizeof(*lpc));
446 for (i = 0; i < SUBFRAMES; i++) {
448 lpc_ptr += LPC_ORDER;
453 * Generate a train of dirac functions with period as pitch lag.
455 static void gen_dirac_train(int16_t *buf, int pitch_lag)
457 int16_t vector[SUBFRAME_LEN];
460 memcpy(vector, buf, SUBFRAME_LEN * sizeof(*vector));
461 for (i = pitch_lag; i < SUBFRAME_LEN; i += pitch_lag) {
462 for (j = 0; j < SUBFRAME_LEN - i; j++)
463 buf[i + j] += vector[j];
468 * Generate fixed codebook excitation vector.
470 * @param vector decoded excitation vector
471 * @param subfrm current subframe
472 * @param cur_rate current bitrate
473 * @param pitch_lag closed loop pitch lag
474 * @param index current subframe index
476 static void gen_fcb_excitation(int16_t *vector, G723_1_Subframe subfrm,
477 enum Rate cur_rate, int pitch_lag, int index)
481 memset(vector, 0, SUBFRAME_LEN * sizeof(*vector));
483 if (cur_rate == RATE_6300) {
484 if (subfrm.pulse_pos >= max_pos[index])
487 /* Decode amplitudes and positions */
488 j = PULSE_MAX - pulses[index];
489 temp = subfrm.pulse_pos;
490 for (i = 0; i < SUBFRAME_LEN / GRID_SIZE; i++) {
491 temp -= combinatorial_table[j][i];
494 temp += combinatorial_table[j++][i];
495 if (subfrm.pulse_sign & (1 << (PULSE_MAX - j))) {
496 vector[subfrm.grid_index + GRID_SIZE * i] =
497 -fixed_cb_gain[subfrm.amp_index];
499 vector[subfrm.grid_index + GRID_SIZE * i] =
500 fixed_cb_gain[subfrm.amp_index];
505 if (subfrm.dirac_train == 1)
506 gen_dirac_train(vector, pitch_lag);
507 } else { /* 5300 bps */
508 int cb_gain = fixed_cb_gain[subfrm.amp_index];
509 int cb_shift = subfrm.grid_index;
510 int cb_sign = subfrm.pulse_sign;
511 int cb_pos = subfrm.pulse_pos;
512 int offset, beta, lag;
514 for (i = 0; i < 8; i += 2) {
515 offset = ((cb_pos & 7) << 3) + cb_shift + i;
516 vector[offset] = (cb_sign & 1) ? cb_gain : -cb_gain;
521 /* Enhance harmonic components */
522 lag = pitch_contrib[subfrm.ad_cb_gain << 1] + pitch_lag +
523 subfrm.ad_cb_lag - 1;
524 beta = pitch_contrib[(subfrm.ad_cb_gain << 1) + 1];
526 if (lag < SUBFRAME_LEN - 2) {
527 for (i = lag; i < SUBFRAME_LEN; i++)
528 vector[i] += beta * vector[i - lag] >> 15;
534 * Get delayed contribution from the previous excitation vector.
536 static void get_residual(int16_t *residual, int16_t *prev_excitation, int lag)
538 int offset = PITCH_MAX - PITCH_ORDER / 2 - lag;
541 residual[0] = prev_excitation[offset];
542 residual[1] = prev_excitation[offset + 1];
545 for (i = 2; i < SUBFRAME_LEN + PITCH_ORDER - 1; i++)
546 residual[i] = prev_excitation[offset + (i - 2) % lag];
550 * Generate adaptive codebook excitation.
552 static void gen_acb_excitation(int16_t *vector, int16_t *prev_excitation,
553 int pitch_lag, G723_1_Subframe subfrm,
556 int16_t residual[SUBFRAME_LEN + PITCH_ORDER - 1];
557 const int16_t *cb_ptr;
558 int lag = pitch_lag + subfrm.ad_cb_lag - 1;
563 get_residual(residual, prev_excitation, lag);
565 /* Select quantization table */
566 if (cur_rate == RATE_6300 && pitch_lag < SUBFRAME_LEN - 2) {
567 cb_ptr = adaptive_cb_gain85;
569 cb_ptr = adaptive_cb_gain170;
571 /* Calculate adaptive vector */
572 cb_ptr += subfrm.ad_cb_gain * 20;
573 for (i = 0; i < SUBFRAME_LEN; i++) {
574 sum = ff_dot_product(residual + i, cb_ptr, PITCH_ORDER);
575 vector[i] = av_clipl_int32((sum << 2) + (1 << 15)) >> 16;
580 * Estimate maximum auto-correlation around pitch lag.
582 * @param p the context
583 * @param offset offset of the excitation vector
584 * @param ccr_max pointer to the maximum auto-correlation
585 * @param pitch_lag decoded pitch lag
586 * @param length length of autocorrelation
587 * @param dir forward lag(1) / backward lag(-1)
589 static int autocorr_max(G723_1_Context *p, int offset, int *ccr_max,
590 int pitch_lag, int length, int dir)
592 int limit, ccr, lag = 0;
593 int16_t *buf = p->excitation + offset;
596 pitch_lag = FFMIN(PITCH_MAX - 3, pitch_lag);
598 limit = FFMIN(FRAME_LEN + PITCH_MAX - offset - length, pitch_lag + 3);
600 limit = pitch_lag + 3;
602 for (i = pitch_lag - 3; i <= limit; i++) {
603 ccr = ff_dot_product(buf, buf + dir * i, length)<<1;
605 if (ccr > *ccr_max) {
614 * Calculate pitch postfilter optimal and scaling gains.
616 * @param lag pitch postfilter forward/backward lag
617 * @param ppf pitch postfilter parameters
618 * @param cur_rate current bitrate
619 * @param tgt_eng target energy
620 * @param ccr cross-correlation
621 * @param res_eng residual energy
623 static void comp_ppf_gains(int lag, PPFParam *ppf, enum Rate cur_rate,
624 int tgt_eng, int ccr, int res_eng)
626 int pf_residual; /* square of postfiltered residual */
627 int64_t temp1, temp2;
631 temp1 = tgt_eng * res_eng >> 1;
632 temp2 = ccr * ccr << 1;
635 if (ccr >= res_eng) {
636 ppf->opt_gain = ppf_gain_weight[cur_rate];
638 ppf->opt_gain = (ccr << 15) / res_eng *
639 ppf_gain_weight[cur_rate] >> 15;
641 /* pf_res^2 = tgt_eng + 2*ccr*gain + res_eng*gain^2 */
642 temp1 = (tgt_eng << 15) + (ccr * ppf->opt_gain << 1);
643 temp2 = (ppf->opt_gain * ppf->opt_gain >> 15) * res_eng;
644 pf_residual = av_clipl_int32(temp1 + temp2 + (1 << 15)) >> 16;
646 if (tgt_eng >= pf_residual << 1) {
649 temp1 = (tgt_eng << 14) / pf_residual;
652 /* scaling_gain = sqrt(tgt_eng/pf_res^2) */
653 ppf->sc_gain = square_root(temp1 << 16);
656 ppf->sc_gain = 0x7fff;
659 ppf->opt_gain = av_clip_int16(ppf->opt_gain * ppf->sc_gain >> 15);
663 * Calculate pitch postfilter parameters.
665 * @param p the context
666 * @param offset offset of the excitation vector
667 * @param pitch_lag decoded pitch lag
668 * @param ppf pitch postfilter parameters
669 * @param cur_rate current bitrate
671 static void comp_ppf_coeff(G723_1_Context *p, int offset, int pitch_lag,
672 PPFParam *ppf, enum Rate cur_rate)
677 int64_t temp1, temp2;
681 * 1 - forward cross-correlation
682 * 2 - forward residual energy
683 * 3 - backward cross-correlation
684 * 4 - backward residual energy
686 int energy[5] = {0, 0, 0, 0, 0};
687 int16_t *buf = p->excitation + offset;
688 int fwd_lag = autocorr_max(p, offset, &energy[1], pitch_lag,
690 int back_lag = autocorr_max(p, offset, &energy[3], pitch_lag,
695 ppf->sc_gain = 0x7fff;
697 /* Case 0, Section 3.6 */
698 if (!back_lag && !fwd_lag)
701 /* Compute target energy */
702 energy[0] = ff_dot_product(buf, buf, SUBFRAME_LEN)<<1;
704 /* Compute forward residual energy */
706 energy[2] = ff_dot_product(buf + fwd_lag, buf + fwd_lag,
709 /* Compute backward residual energy */
711 energy[4] = ff_dot_product(buf - back_lag, buf - back_lag,
714 /* Normalize and shorten */
716 for (i = 0; i < 5; i++)
717 temp1 = FFMAX(energy[i], temp1);
719 scale = normalize_bits(temp1, 31);
720 for (i = 0; i < 5; i++)
721 energy[i] = av_clipl_int32(energy[i] << scale) >> 16;
723 if (fwd_lag && !back_lag) { /* Case 1 */
724 comp_ppf_gains(fwd_lag, ppf, cur_rate, energy[0], energy[1],
726 } else if (!fwd_lag) { /* Case 2 */
727 comp_ppf_gains(-back_lag, ppf, cur_rate, energy[0], energy[3],
729 } else { /* Case 3 */
732 * Select the largest of energy[1]^2/energy[2]
733 * and energy[3]^2/energy[4]
735 temp1 = energy[4] * ((energy[1] * energy[1] + (1 << 14)) >> 15);
736 temp2 = energy[2] * ((energy[3] * energy[3] + (1 << 14)) >> 15);
737 if (temp1 >= temp2) {
738 comp_ppf_gains(fwd_lag, ppf, cur_rate, energy[0], energy[1],
741 comp_ppf_gains(-back_lag, ppf, cur_rate, energy[0], energy[3],
748 * Classify frames as voiced/unvoiced.
750 * @param p the context
751 * @param pitch_lag decoded pitch_lag
752 * @param exc_eng excitation energy estimation
753 * @param scale scaling factor of exc_eng
755 * @return residual interpolation index if voiced, 0 otherwise
757 static int comp_interp_index(G723_1_Context *p, int pitch_lag,
758 int *exc_eng, int *scale)
760 int offset = PITCH_MAX + 2 * SUBFRAME_LEN;
761 int16_t *buf = p->excitation + offset;
763 int index, ccr, tgt_eng, best_eng, temp;
765 *scale = scale_vector(p->excitation, FRAME_LEN + PITCH_MAX);
767 /* Compute maximum backward cross-correlation */
769 index = autocorr_max(p, offset, &ccr, pitch_lag, SUBFRAME_LEN * 2, -1);
770 ccr = av_clipl_int32((int64_t)ccr + (1 << 15)) >> 16;
772 /* Compute target energy */
773 tgt_eng = ff_dot_product(buf, buf, SUBFRAME_LEN * 2)<<1;
774 *exc_eng = av_clipl_int32(tgt_eng + (1 << 15)) >> 16;
779 /* Compute best energy */
780 best_eng = ff_dot_product(buf - index, buf - index,
781 SUBFRAME_LEN * 2)<<1;
782 best_eng = av_clipl_int32((int64_t)best_eng + (1 << 15)) >> 16;
784 temp = best_eng * *exc_eng >> 3;
786 if (temp < ccr * ccr) {
793 * Peform residual interpolation based on frame classification.
795 * @param buf decoded excitation vector
796 * @param out output vector
797 * @param lag decoded pitch lag
798 * @param gain interpolated gain
799 * @param rseed seed for random number generator
801 static void residual_interp(int16_t *buf, int16_t *out, int lag,
802 int gain, int *rseed)
805 if (lag) { /* Voiced */
806 int16_t *vector_ptr = buf + PITCH_MAX;
808 for (i = 0; i < lag; i++)
809 vector_ptr[i - lag] = vector_ptr[i - lag] * 3 >> 2;
810 av_memcpy_backptr((uint8_t*)vector_ptr, lag * sizeof(*vector_ptr),
811 FRAME_LEN * sizeof(*vector_ptr));
812 memcpy(out, vector_ptr, FRAME_LEN * sizeof(*vector_ptr));
813 } else { /* Unvoiced */
814 for (i = 0; i < FRAME_LEN; i++) {
815 *rseed = *rseed * 521 + 259;
816 out[i] = gain * *rseed >> 15;
818 memset(buf, 0, (FRAME_LEN + PITCH_MAX) * sizeof(*buf));
823 * Perform IIR filtering.
825 * @param fir_coef FIR coefficients
826 * @param iir_coef IIR coefficients
827 * @param src source vector
828 * @param dest destination vector
829 * @param width width of the output, 16 bits(0) / 32 bits(1)
831 #define iir_filter(fir_coef, iir_coef, src, dest, width)\
834 int res_shift = 16 & ~-(width);\
835 int in_shift = 16 - res_shift;\
837 for (m = 0; m < SUBFRAME_LEN; m++) {\
839 for (n = 1; n <= LPC_ORDER; n++) {\
840 filter -= (fir_coef)[n - 1] * (src)[m - n] -\
841 (iir_coef)[n - 1] * ((dest)[m - n] >> in_shift);\
844 (dest)[m] = av_clipl_int32(((src)[m] << 16) + (filter << 3) +\
845 (1 << 15)) >> res_shift;\
850 * Adjust gain of postfiltered signal.
852 * @param p the context
853 * @param buf postfiltered output vector
854 * @param energy input energy coefficient
856 static void gain_scale(G723_1_Context *p, int16_t * buf, int energy)
858 int num, denom, gain, bits1, bits2;
863 for (i = 0; i < SUBFRAME_LEN; i++) {
864 int64_t temp = buf[i] >> 2;
865 temp = av_clipl_int32(MUL64(temp, temp) << 1);
866 denom = av_clipl_int32(denom + temp);
870 bits1 = normalize_bits(num, 31);
871 bits2 = normalize_bits(denom, 31);
872 num = num << bits1 >> 1;
875 bits2 = 5 + bits1 - bits2;
876 bits2 = FFMAX(0, bits2);
878 gain = (num >> 1) / (denom >> 16);
879 gain = square_root(gain << 16 >> bits2);
884 for (i = 0; i < SUBFRAME_LEN; i++) {
885 p->pf_gain = ((p->pf_gain << 4) - p->pf_gain + gain + (1 << 3)) >> 4;
886 buf[i] = av_clip_int16((buf[i] * (p->pf_gain + (p->pf_gain >> 4)) +
892 * Perform formant filtering.
894 * @param p the context
895 * @param lpc quantized lpc coefficients
896 * @param buf output buffer
898 static void formant_postfilter(G723_1_Context *p, int16_t *lpc, int16_t *buf)
900 int16_t filter_coef[2][LPC_ORDER], *buf_ptr;
901 int filter_signal[LPC_ORDER + FRAME_LEN], *signal_ptr;
904 memcpy(buf, p->fir_mem, LPC_ORDER * sizeof(*buf));
905 memcpy(filter_signal, p->iir_mem, LPC_ORDER * sizeof(*filter_signal));
907 for (i = LPC_ORDER, j = 0; j < SUBFRAMES; i += SUBFRAME_LEN, j++) {
908 for (k = 0; k < LPC_ORDER; k++) {
909 filter_coef[0][k] = (-lpc[k] * postfilter_tbl[0][k] +
911 filter_coef[1][k] = (-lpc[k] * postfilter_tbl[1][k] +
914 iir_filter(filter_coef[0], filter_coef[1], buf + i,
915 filter_signal + i, 1);
919 memcpy(p->fir_mem, buf + FRAME_LEN, LPC_ORDER * sizeof(int16_t));
920 memcpy(p->iir_mem, filter_signal + FRAME_LEN, LPC_ORDER * sizeof(int));
922 buf_ptr = buf + LPC_ORDER;
923 signal_ptr = filter_signal + LPC_ORDER;
924 for (i = 0; i < SUBFRAMES; i++) {
925 int16_t temp_vector[SUBFRAME_LEN];
931 memcpy(temp_vector, buf_ptr, SUBFRAME_LEN * sizeof(*temp_vector));
932 scale = scale_vector(temp_vector, SUBFRAME_LEN);
934 /* Compute auto correlation coefficients */
935 auto_corr[0] = ff_dot_product(temp_vector, temp_vector + 1,
936 SUBFRAME_LEN - 1)<<1;
937 auto_corr[1] = ff_dot_product(temp_vector, temp_vector,
940 /* Compute reflection coefficient */
941 temp = auto_corr[1] >> 16;
943 temp = (auto_corr[0] >> 2) / temp;
945 p->reflection_coef = (3 * p->reflection_coef + temp + 2) >> 2;
946 temp = -p->reflection_coef >> 1 & ~3;
948 /* Compensation filter */
949 for (j = 0; j < SUBFRAME_LEN; j++) {
950 buf_ptr[j] = av_clipl_int32((int64_t)signal_ptr[j] +
951 ((signal_ptr[j - 1] >> 16) *
955 /* Compute normalized signal energy */
956 temp = 2 * scale + 4;
958 energy = av_clipl_int32((int64_t)auto_corr[1] << -temp);
960 energy = auto_corr[1] >> temp;
962 gain_scale(p, buf_ptr, energy);
964 buf_ptr += SUBFRAME_LEN;
965 signal_ptr += SUBFRAME_LEN;
969 static int g723_1_decode_frame(AVCodecContext *avctx, void *data,
970 int *got_frame_ptr, AVPacket *avpkt)
972 G723_1_Context *p = avctx->priv_data;
973 const uint8_t *buf = avpkt->data;
974 int buf_size = avpkt->size;
975 int dec_mode = buf[0] & 3;
977 PPFParam ppf[SUBFRAMES];
978 int16_t cur_lsp[LPC_ORDER];
979 int16_t lpc[SUBFRAMES * LPC_ORDER];
980 int16_t acb_vector[SUBFRAME_LEN];
983 int bad_frame = 0, i, j, ret;
985 if (buf_size < frame_size[dec_mode]) {
987 av_log(avctx, AV_LOG_WARNING,
988 "Expected %d bytes, got %d - skipping packet\n",
989 frame_size[dec_mode], buf_size);
994 if (unpack_bitstream(p, buf, buf_size) < 0) {
996 if (p->past_frame_type == ACTIVE_FRAME)
997 p->cur_frame_type = ACTIVE_FRAME;
999 p->cur_frame_type = UNTRANSMITTED_FRAME;
1002 p->frame.nb_samples = FRAME_LEN;
1003 if ((ret = avctx->get_buffer(avctx, &p->frame)) < 0) {
1004 av_log(avctx, AV_LOG_ERROR, "get_buffer() failed\n");
1008 out = (int16_t *)p->frame.data[0];
1010 if (p->cur_frame_type == ACTIVE_FRAME) {
1012 p->erased_frames = 0;
1013 else if (p->erased_frames != 3)
1016 inverse_quant(cur_lsp, p->prev_lsp, p->lsp_index, bad_frame);
1017 lsp_interpolate(lpc, cur_lsp, p->prev_lsp);
1019 /* Save the lsp_vector for the next frame */
1020 memcpy(p->prev_lsp, cur_lsp, LPC_ORDER * sizeof(*p->prev_lsp));
1022 /* Generate the excitation for the frame */
1023 memcpy(p->excitation, p->prev_excitation,
1024 PITCH_MAX * sizeof(*p->excitation));
1025 vector_ptr = p->excitation + PITCH_MAX;
1026 if (!p->erased_frames) {
1027 /* Update interpolation gain memory */
1028 p->interp_gain = fixed_cb_gain[(p->subframe[2].amp_index +
1029 p->subframe[3].amp_index) >> 1];
1030 for (i = 0; i < SUBFRAMES; i++) {
1031 gen_fcb_excitation(vector_ptr, p->subframe[i], p->cur_rate,
1032 p->pitch_lag[i >> 1], i);
1033 gen_acb_excitation(acb_vector, &p->excitation[SUBFRAME_LEN * i],
1034 p->pitch_lag[i >> 1], p->subframe[i],
1036 /* Get the total excitation */
1037 for (j = 0; j < SUBFRAME_LEN; j++) {
1038 vector_ptr[j] = av_clip_int16(vector_ptr[j] << 1);
1039 vector_ptr[j] = av_clip_int16(vector_ptr[j] +
1042 vector_ptr += SUBFRAME_LEN;
1045 vector_ptr = p->excitation + PITCH_MAX;
1047 /* Save the excitation */
1048 memcpy(p->audio + LPC_ORDER, vector_ptr, FRAME_LEN * sizeof(*p->audio));
1050 p->interp_index = comp_interp_index(p, p->pitch_lag[1],
1051 &p->sid_gain, &p->cur_gain);
1053 if (p->postfilter) {
1055 for (j = 0; j < SUBFRAMES; i += SUBFRAME_LEN, j++)
1056 comp_ppf_coeff(p, i, p->pitch_lag[j >> 1],
1057 ppf + j, p->cur_rate);
1060 /* Restore the original excitation */
1061 memcpy(p->excitation, p->prev_excitation,
1062 PITCH_MAX * sizeof(*p->excitation));
1063 memcpy(vector_ptr, p->audio + LPC_ORDER, FRAME_LEN * sizeof(*vector_ptr));
1065 /* Peform pitch postfiltering */
1067 for (i = 0, j = 0; j < SUBFRAMES; i += SUBFRAME_LEN, j++)
1068 ff_acelp_weighted_vector_sum(p->audio + LPC_ORDER + i,
1070 vector_ptr + i + ppf[j].index,
1073 1 << 14, 15, SUBFRAME_LEN);
1076 p->interp_gain = (p->interp_gain * 3 + 2) >> 2;
1077 if (p->erased_frames == 3) {
1079 memset(p->excitation, 0,
1080 (FRAME_LEN + PITCH_MAX) * sizeof(*p->excitation));
1081 memset(p->frame.data[0], 0,
1082 (FRAME_LEN + LPC_ORDER) * sizeof(int16_t));
1084 /* Regenerate frame */
1085 residual_interp(p->excitation, p->audio + LPC_ORDER, p->interp_index,
1086 p->interp_gain, &p->random_seed);
1089 /* Save the excitation for the next frame */
1090 memcpy(p->prev_excitation, p->excitation + FRAME_LEN,
1091 PITCH_MAX * sizeof(*p->excitation));
1093 memset(out, 0, FRAME_LEN * 2);
1094 av_log(avctx, AV_LOG_WARNING,
1095 "G.723.1: Comfort noise generation not supported yet\n");
1098 *(AVFrame *)data = p->frame;
1099 return frame_size[dec_mode];
1102 p->past_frame_type = p->cur_frame_type;
1104 memcpy(p->audio, p->synth_mem, LPC_ORDER * sizeof(*p->audio));
1105 for (i = LPC_ORDER, j = 0; j < SUBFRAMES; i += SUBFRAME_LEN, j++)
1106 ff_celp_lp_synthesis_filter(p->audio + i, &lpc[j * LPC_ORDER],
1107 p->audio + i, SUBFRAME_LEN, LPC_ORDER,
1109 memcpy(p->synth_mem, p->audio + FRAME_LEN, LPC_ORDER * sizeof(*p->audio));
1111 if (p->postfilter) {
1112 formant_postfilter(p, lpc, p->audio);
1113 memcpy(p->frame.data[0], p->audio + LPC_ORDER, FRAME_LEN * 2);
1114 } else { // if output is not postfiltered it should be scaled by 2
1115 for (i = 0; i < FRAME_LEN; i++)
1116 out[i] = av_clip_int16(p->audio[LPC_ORDER + i] << 1);
1120 *(AVFrame *)data = p->frame;
1122 return frame_size[dec_mode];
1125 #define OFFSET(x) offsetof(G723_1_Context, x)
1126 #define AD AV_OPT_FLAG_AUDIO_PARAM | AV_OPT_FLAG_DECODING_PARAM
1128 static const AVOption options[] = {
1129 { "postfilter", "postfilter on/off", OFFSET(postfilter), AV_OPT_TYPE_INT,
1135 static const AVClass g723_1dec_class = {
1136 .class_name = "G.723.1 decoder",
1137 .item_name = av_default_item_name,
1139 .version = LIBAVUTIL_VERSION_INT,
1142 AVCodec ff_g723_1_decoder = {
1144 .type = AVMEDIA_TYPE_AUDIO,
1145 .id = AV_CODEC_ID_G723_1,
1146 .priv_data_size = sizeof(G723_1_Context),
1147 .init = g723_1_decode_init,
1148 .decode = g723_1_decode_frame,
1149 .long_name = NULL_IF_CONFIG_SMALL("G.723.1"),
1150 .capabilities = CODEC_CAP_SUBFRAMES | CODEC_CAP_DR1,
1151 .priv_class = &g723_1dec_class,
1154 #if CONFIG_G723_1_ENCODER
1155 #define BITSTREAM_WRITER_LE
1156 #include "put_bits.h"
1158 static av_cold int g723_1_encode_init(AVCodecContext *avctx)
1160 G723_1_Context *p = avctx->priv_data;
1162 if (avctx->sample_rate != 8000) {
1163 av_log(avctx, AV_LOG_ERROR, "Only 8000Hz sample rate supported\n");
1167 if (avctx->channels != 1) {
1168 av_log(avctx, AV_LOG_ERROR, "Only mono supported\n");
1169 return AVERROR(EINVAL);
1172 if (avctx->bit_rate == 6300) {
1173 p->cur_rate = RATE_6300;
1174 } else if (avctx->bit_rate == 5300) {
1175 av_log(avctx, AV_LOG_ERROR, "Bitrate not supported yet, use 6.3k\n");
1176 return AVERROR_PATCHWELCOME;
1178 av_log(avctx, AV_LOG_ERROR,
1179 "Bitrate not supported, use 6.3k\n");
1180 return AVERROR(EINVAL);
1182 avctx->frame_size = 240;
1183 memcpy(p->prev_lsp, dc_lsp, LPC_ORDER * sizeof(int16_t));
1189 * Remove DC component from the input signal.
1191 * @param buf input signal
1192 * @param fir zero memory
1193 * @param iir pole memory
1195 static void highpass_filter(int16_t *buf, int16_t *fir, int *iir)
1198 for (i = 0; i < FRAME_LEN; i++) {
1199 *iir = (buf[i] << 15) + ((-*fir) << 15) + MULL2(*iir, 0x7f00);
1201 buf[i] = av_clipl_int32((int64_t)*iir + (1 << 15)) >> 16;
1206 * Estimate autocorrelation of the input vector.
1208 * @param buf input buffer
1209 * @param autocorr autocorrelation coefficients vector
1211 static void comp_autocorr(int16_t *buf, int16_t *autocorr)
1214 int16_t vector[LPC_FRAME];
1216 memcpy(vector, buf, LPC_FRAME * sizeof(int16_t));
1217 scale_vector(vector, LPC_FRAME);
1219 /* Apply the Hamming window */
1220 for (i = 0; i < LPC_FRAME; i++)
1221 vector[i] = (vector[i] * hamming_window[i] + (1 << 14)) >> 15;
1223 /* Compute the first autocorrelation coefficient */
1224 temp = dot_product(vector, vector, LPC_FRAME, 0);
1226 /* Apply a white noise correlation factor of (1025/1024) */
1230 scale = normalize_bits_int32(temp);
1231 autocorr[0] = av_clipl_int32((int64_t)(temp << scale) +
1234 /* Compute the remaining coefficients */
1236 memset(autocorr + 1, 0, LPC_ORDER * sizeof(int16_t));
1238 for (i = 1; i <= LPC_ORDER; i++) {
1239 temp = dot_product(vector, vector + i, LPC_FRAME - i, 0);
1240 temp = MULL2((temp << scale), binomial_window[i - 1]);
1241 autocorr[i] = av_clipl_int32((int64_t)temp + (1 << 15)) >> 16;
1247 * Use Levinson-Durbin recursion to compute LPC coefficients from
1248 * autocorrelation values.
1250 * @param lpc LPC coefficients vector
1251 * @param autocorr autocorrelation coefficients vector
1252 * @param error prediction error
1254 static void levinson_durbin(int16_t *lpc, int16_t *autocorr, int16_t error)
1256 int16_t vector[LPC_ORDER];
1257 int16_t partial_corr;
1260 memset(lpc, 0, LPC_ORDER * sizeof(int16_t));
1262 for (i = 0; i < LPC_ORDER; i++) {
1263 /* Compute the partial correlation coefficient */
1265 for (j = 0; j < i; j++)
1266 temp -= lpc[j] * autocorr[i - j - 1];
1267 temp = ((autocorr[i] << 13) + temp) << 3;
1269 if (FFABS(temp) >= (error << 16))
1272 partial_corr = temp / (error << 1);
1274 lpc[i] = av_clipl_int32((int64_t)(partial_corr << 14) +
1277 /* Update the prediction error */
1278 temp = MULL2(temp, partial_corr);
1279 error = av_clipl_int32((int64_t)(error << 16) - temp +
1282 memcpy(vector, lpc, i * sizeof(int16_t));
1283 for (j = 0; j < i; j++) {
1284 temp = partial_corr * vector[i - j - 1] << 1;
1285 lpc[j] = av_clipl_int32((int64_t)(lpc[j] << 16) - temp +
1292 * Calculate LPC coefficients for the current frame.
1294 * @param buf current frame
1295 * @param prev_data 2 trailing subframes of the previous frame
1296 * @param lpc LPC coefficients vector
1298 static void comp_lpc_coeff(int16_t *buf, int16_t *lpc)
1300 int16_t autocorr[(LPC_ORDER + 1) * SUBFRAMES];
1301 int16_t *autocorr_ptr = autocorr;
1302 int16_t *lpc_ptr = lpc;
1305 for (i = 0, j = 0; j < SUBFRAMES; i += SUBFRAME_LEN, j++) {
1306 comp_autocorr(buf + i, autocorr_ptr);
1307 levinson_durbin(lpc_ptr, autocorr_ptr + 1, autocorr_ptr[0]);
1309 lpc_ptr += LPC_ORDER;
1310 autocorr_ptr += LPC_ORDER + 1;
1314 static void lpc2lsp(int16_t *lpc, int16_t *prev_lsp, int16_t *lsp)
1316 int f[LPC_ORDER + 2]; ///< coefficients of the sum and difference
1317 ///< polynomials (F1, F2) ordered as
1318 ///< f1[0], f2[0], ...., f1[5], f2[5]
1320 int max, shift, cur_val, prev_val, count, p;
1324 /* Initialize f1[0] and f2[0] to 1 in Q25 */
1325 for (i = 0; i < LPC_ORDER; i++)
1326 lsp[i] = (lpc[i] * bandwidth_expand[i] + (1 << 14)) >> 15;
1328 /* Apply bandwidth expansion on the LPC coefficients */
1329 f[0] = f[1] = 1 << 25;
1331 /* Compute the remaining coefficients */
1332 for (i = 0; i < LPC_ORDER / 2; i++) {
1334 f[2 * i + 2] = -f[2 * i] - ((lsp[i] + lsp[LPC_ORDER - 1 - i]) << 12);
1336 f[2 * i + 3] = f[2 * i + 1] - ((lsp[i] - lsp[LPC_ORDER - 1 - i]) << 12);
1339 /* Divide f1[5] and f2[5] by 2 for use in polynomial evaluation */
1341 f[LPC_ORDER + 1] >>= 1;
1343 /* Normalize and shorten */
1345 for (i = 1; i < LPC_ORDER + 2; i++)
1346 max = FFMAX(max, FFABS(f[i]));
1348 shift = normalize_bits_int32(max);
1350 for (i = 0; i < LPC_ORDER + 2; i++)
1351 f[i] = av_clipl_int32((int64_t)(f[i] << shift) + (1 << 15)) >> 16;
1354 * Evaluate F1 and F2 at uniform intervals of pi/256 along the
1355 * unit circle and check for zero crossings.
1359 for (i = 0; i <= LPC_ORDER / 2; i++)
1360 temp += f[2 * i] * cos_tab[0];
1361 prev_val = av_clipl_int32(temp << 1);
1363 for ( i = 1; i < COS_TBL_SIZE / 2; i++) {
1366 for (j = 0; j <= LPC_ORDER / 2; j++)
1367 temp += f[LPC_ORDER - 2 * j + p] * cos_tab[i * j % COS_TBL_SIZE];
1368 cur_val = av_clipl_int32(temp << 1);
1370 /* Check for sign change, indicating a zero crossing */
1371 if ((cur_val ^ prev_val) < 0) {
1372 int abs_cur = FFABS(cur_val);
1373 int abs_prev = FFABS(prev_val);
1374 int sum = abs_cur + abs_prev;
1376 shift = normalize_bits_int32(sum);
1378 abs_prev = abs_prev << shift >> 8;
1379 lsp[count++] = ((i - 1) << 7) + (abs_prev >> 1) / (sum >> 16);
1381 if (count == LPC_ORDER)
1384 /* Switch between sum and difference polynomials */
1389 for (j = 0; j <= LPC_ORDER / 2; j++){
1390 temp += f[LPC_ORDER - 2 * j + p] *
1391 cos_tab[i * j % COS_TBL_SIZE];
1393 cur_val = av_clipl_int32(temp<<1);
1398 if (count != LPC_ORDER)
1399 memcpy(lsp, prev_lsp, LPC_ORDER * sizeof(int16_t));
1403 * Quantize the current LSP subvector.
1405 * @param num band number
1406 * @param offset offset of the current subvector in an LPC_ORDER vector
1407 * @param size size of the current subvector
1409 #define get_index(num, offset, size) \
1411 int error, max = -1;\
1414 for (i = 0; i < LSP_CB_SIZE; i++) {\
1415 for (j = 0; j < size; j++){\
1416 temp[j] = (weight[j + (offset)] * lsp_band##num[i][j] +\
1419 error = dot_product(lsp + (offset), temp, size, 1) << 1;\
1420 error -= dot_product(lsp_band##num[i], temp, size, 1);\
1423 lsp_index[num] = i;\
1429 * Vector quantize the LSP frequencies.
1431 * @param lsp the current lsp vector
1432 * @param prev_lsp the previous lsp vector
1434 static void lsp_quantize(uint8_t *lsp_index, int16_t *lsp, int16_t *prev_lsp)
1436 int16_t weight[LPC_ORDER];
1440 /* Calculate the VQ weighting vector */
1441 weight[0] = (1 << 20) / (lsp[1] - lsp[0]);
1442 weight[LPC_ORDER - 1] = (1 << 20) /
1443 (lsp[LPC_ORDER - 1] - lsp[LPC_ORDER - 2]);
1445 for (i = 1; i < LPC_ORDER - 1; i++) {
1446 min = FFMIN(lsp[i] - lsp[i - 1], lsp[i + 1] - lsp[i]);
1448 weight[i] = (1 << 20) / min;
1450 weight[i] = INT16_MAX;
1455 for (i = 0; i < LPC_ORDER; i++)
1456 max = FFMAX(weight[i], max);
1458 shift = normalize_bits_int16(max);
1459 for (i = 0; i < LPC_ORDER; i++) {
1460 weight[i] <<= shift;
1463 /* Compute the VQ target vector */
1464 for (i = 0; i < LPC_ORDER; i++) {
1465 lsp[i] -= dc_lsp[i] +
1466 (((prev_lsp[i] - dc_lsp[i]) * 12288 + (1 << 14)) >> 15);
1475 * Apply the formant perceptual weighting filter.
1477 * @param flt_coef filter coefficients
1478 * @param unq_lpc unquantized lpc vector
1480 static void perceptual_filter(G723_1_Context *p, int16_t *flt_coef,
1481 int16_t *unq_lpc, int16_t *buf)
1483 int16_t vector[FRAME_LEN + LPC_ORDER];
1486 memcpy(buf, p->iir_mem, sizeof(int16_t) * LPC_ORDER);
1487 memcpy(vector, p->fir_mem, sizeof(int16_t) * LPC_ORDER);
1488 memcpy(vector + LPC_ORDER, buf + LPC_ORDER, sizeof(int16_t) * FRAME_LEN);
1490 for (i = LPC_ORDER, j = 0; j < SUBFRAMES; i += SUBFRAME_LEN, j++) {
1491 for (k = 0; k < LPC_ORDER; k++) {
1492 flt_coef[k + 2 * l] = (unq_lpc[k + l] * percept_flt_tbl[0][k] +
1494 flt_coef[k + 2 * l + LPC_ORDER] = (unq_lpc[k + l] *
1495 percept_flt_tbl[1][k] +
1498 iir_filter(flt_coef + 2 * l, flt_coef + 2 * l + LPC_ORDER, vector + i,
1502 memcpy(p->iir_mem, buf + FRAME_LEN, sizeof(int16_t) * LPC_ORDER);
1503 memcpy(p->fir_mem, vector + FRAME_LEN, sizeof(int16_t) * LPC_ORDER);
1507 * Estimate the open loop pitch period.
1509 * @param buf perceptually weighted speech
1510 * @param start estimation is carried out from this position
1512 static int estimate_pitch(int16_t *buf, int start)
1515 int max_ccr = 0x4000;
1516 int max_eng = 0x7fff;
1517 int index = PITCH_MIN;
1518 int offset = start - PITCH_MIN + 1;
1520 int ccr, eng, orig_eng, ccr_eng, exp;
1525 orig_eng = dot_product(buf + offset, buf + offset, HALF_FRAME_LEN, 0);
1527 for (i = PITCH_MIN; i <= PITCH_MAX - 3; i++) {
1530 /* Update energy and compute correlation */
1531 orig_eng += buf[offset] * buf[offset] -
1532 buf[offset + HALF_FRAME_LEN] * buf[offset + HALF_FRAME_LEN];
1533 ccr = dot_product(buf + start, buf + offset, HALF_FRAME_LEN, 0);
1537 /* Split into mantissa and exponent to maintain precision */
1538 exp = normalize_bits_int32(ccr);
1539 ccr = av_clipl_int32((int64_t)(ccr << exp) + (1 << 15)) >> 16;
1542 temp = normalize_bits_int32(ccr);
1543 ccr = ccr << temp >> 16;
1546 temp = normalize_bits_int32(orig_eng);
1547 eng = av_clipl_int32((int64_t)(orig_eng << temp) + (1 << 15)) >> 16;
1557 if (exp + 1 < max_exp)
1560 /* Equalize exponents before comparison */
1561 if (exp + 1 == max_exp)
1562 temp = max_ccr >> 1;
1565 ccr_eng = ccr * max_eng;
1566 diff = ccr_eng - eng * temp;
1567 if (diff > 0 && (i - index < PITCH_MIN || diff > ccr_eng >> 2)) {
1579 * Compute harmonic noise filter parameters.
1581 * @param buf perceptually weighted speech
1582 * @param pitch_lag open loop pitch period
1583 * @param hf harmonic filter parameters
1585 static void comp_harmonic_coeff(int16_t *buf, int16_t pitch_lag, HFParam *hf)
1587 int ccr, eng, max_ccr, max_eng;
1592 for (i = 0, j = pitch_lag - 3; j <= pitch_lag + 3; i++, j++) {
1593 /* Compute residual energy */
1594 energy[i << 1] = dot_product(buf - j, buf - j, SUBFRAME_LEN, 0);
1595 /* Compute correlation */
1596 energy[(i << 1) + 1] = dot_product(buf, buf - j, SUBFRAME_LEN, 0);
1599 /* Compute target energy */
1600 energy[14] = dot_product(buf, buf, SUBFRAME_LEN, 0);
1604 for (i = 0; i < 15; i++)
1605 max = FFMAX(max, FFABS(energy[i]));
1607 exp = normalize_bits_int32(max);
1608 for (i = 0; i < 15; i++) {
1609 energy[i] = av_clipl_int32((int64_t)(energy[i] << exp) +
1618 for (i = 0; i <= 6; i++) {
1619 eng = energy[i << 1];
1620 ccr = energy[(i << 1) + 1];
1625 ccr = (ccr * ccr + (1 << 14)) >> 15;
1626 diff = ccr * max_eng - eng * max_ccr;
1634 if (hf->index == -1) {
1635 hf->index = pitch_lag;
1639 eng = energy[14] * max_eng;
1640 eng = (eng >> 2) + (eng >> 3);
1641 ccr = energy[(hf->index << 1) + 1] * energy[(hf->index << 1) + 1];
1643 eng = energy[(hf->index << 1) + 1];
1648 hf->gain = ((eng << 15) / max_eng * 0x2800 + (1 << 14)) >> 15;
1650 hf->index += pitch_lag - 3;
1654 * Apply the harmonic noise shaping filter.
1656 * @param hf filter parameters
1658 static void harmonic_filter(HFParam *hf, int16_t *src, int16_t *dest)
1662 for (i = 0; i < SUBFRAME_LEN; i++) {
1663 int64_t temp = hf->gain * src[i - hf->index] << 1;
1664 dest[i] = av_clipl_int32((src[i] << 16) - temp + (1 << 15)) >> 16;
1668 static void harmonic_noise_sub(HFParam *hf, int16_t *src, int16_t *dest)
1671 for (i = 0; i < SUBFRAME_LEN; i++) {
1672 int64_t temp = hf->gain * src[i - hf->index] << 1;
1673 dest[i] = av_clipl_int32(((dest[i] - src[i]) << 16) + temp +
1680 * Combined synthesis and formant perceptual weighting filer.
1682 * @param qnt_lpc quantized lpc coefficients
1683 * @param perf_lpc perceptual filter coefficients
1684 * @param perf_fir perceptual filter fir memory
1685 * @param perf_iir perceptual filter iir memory
1686 * @param scale the filter output will be scaled by 2^scale
1688 static void synth_percept_filter(int16_t *qnt_lpc, int16_t *perf_lpc,
1689 int16_t *perf_fir, int16_t *perf_iir,
1690 int16_t *src, int16_t *dest, int scale)
1693 int16_t buf_16[SUBFRAME_LEN + LPC_ORDER];
1694 int64_t buf[SUBFRAME_LEN];
1696 int16_t *bptr_16 = buf_16 + LPC_ORDER;
1698 memcpy(buf_16, perf_fir, sizeof(int16_t) * LPC_ORDER);
1699 memcpy(dest - LPC_ORDER, perf_iir, sizeof(int16_t) * LPC_ORDER);
1701 for (i = 0; i < SUBFRAME_LEN; i++) {
1703 for (j = 1; j <= LPC_ORDER; j++)
1704 temp -= qnt_lpc[j - 1] * bptr_16[i - j];
1706 buf[i] = (src[i] << 15) + (temp << 3);
1707 bptr_16[i] = av_clipl_int32(buf[i] + (1 << 15)) >> 16;
1710 for (i = 0; i < SUBFRAME_LEN; i++) {
1711 int64_t fir = 0, iir = 0;
1712 for (j = 1; j <= LPC_ORDER; j++) {
1713 fir -= perf_lpc[j - 1] * bptr_16[i - j];
1714 iir += perf_lpc[j + LPC_ORDER - 1] * dest[i - j];
1716 dest[i] = av_clipl_int32(((buf[i] + (fir << 3)) << scale) + (iir << 3) +
1719 memcpy(perf_fir, buf_16 + SUBFRAME_LEN, sizeof(int16_t) * LPC_ORDER);
1720 memcpy(perf_iir, dest + SUBFRAME_LEN - LPC_ORDER,
1721 sizeof(int16_t) * LPC_ORDER);
1725 * Compute the adaptive codebook contribution.
1727 * @param buf input signal
1728 * @param index the current subframe index
1730 static void acb_search(G723_1_Context *p, int16_t *residual,
1731 int16_t *impulse_resp, int16_t *buf,
1735 int16_t flt_buf[PITCH_ORDER][SUBFRAME_LEN];
1737 const int16_t *cb_tbl = adaptive_cb_gain85;
1739 int ccr_buf[PITCH_ORDER * SUBFRAMES << 2];
1741 int pitch_lag = p->pitch_lag[index >> 1];
1744 int odd_frame = index & 1;
1745 int iter = 3 + odd_frame;
1749 int i, j, k, l, max;
1753 if (pitch_lag == PITCH_MIN)
1756 pitch_lag = FFMIN(pitch_lag, PITCH_MAX - 5);
1759 for (i = 0; i < iter; i++) {
1760 get_residual(residual, p->prev_excitation, pitch_lag + i - 1);
1762 for (j = 0; j < SUBFRAME_LEN; j++) {
1764 for (k = 0; k <= j; k++)
1765 temp += residual[PITCH_ORDER - 1 + k] * impulse_resp[j - k];
1766 flt_buf[PITCH_ORDER - 1][j] = av_clipl_int32((temp << 1) +
1770 for (j = PITCH_ORDER - 2; j >= 0; j--) {
1771 flt_buf[j][0] = ((residual[j] << 13) + (1 << 14)) >> 15;
1772 for (k = 1; k < SUBFRAME_LEN; k++) {
1773 temp = (flt_buf[j + 1][k - 1] << 15) +
1774 residual[j] * impulse_resp[k];
1775 flt_buf[j][k] = av_clipl_int32((temp << 1) + (1 << 15)) >> 16;
1779 /* Compute crosscorrelation with the signal */
1780 for (j = 0; j < PITCH_ORDER; j++) {
1781 temp = dot_product(buf, flt_buf[j], SUBFRAME_LEN, 0);
1782 ccr_buf[count++] = av_clipl_int32(temp << 1);
1785 /* Compute energies */
1786 for (j = 0; j < PITCH_ORDER; j++) {
1787 ccr_buf[count++] = dot_product(flt_buf[j], flt_buf[j],
1791 for (j = 1; j < PITCH_ORDER; j++) {
1792 for (k = 0; k < j; k++) {
1793 temp = dot_product(flt_buf[j], flt_buf[k], SUBFRAME_LEN, 0);
1794 ccr_buf[count++] = av_clipl_int32(temp<<2);
1799 /* Normalize and shorten */
1801 for (i = 0; i < 20 * iter; i++)
1802 max = FFMAX(max, FFABS(ccr_buf[i]));
1804 temp = normalize_bits_int32(max);
1806 for (i = 0; i < 20 * iter; i++){
1807 ccr_buf[i] = av_clipl_int32((int64_t)(ccr_buf[i] << temp) +
1812 for (i = 0; i < iter; i++) {
1813 /* Select quantization table */
1814 if (!odd_frame && pitch_lag + i - 1 >= SUBFRAME_LEN - 2 ||
1815 odd_frame && pitch_lag >= SUBFRAME_LEN - 2) {
1816 cb_tbl = adaptive_cb_gain170;
1820 for (j = 0, k = 0; j < tbl_size; j++, k += 20) {
1822 for (l = 0; l < 20; l++)
1823 temp += ccr_buf[20 * i + l] * cb_tbl[k + l];
1824 temp = av_clipl_int32(temp);
1835 pitch_lag += acb_lag - 1;
1839 p->pitch_lag[index >> 1] = pitch_lag;
1840 p->subframe[index].ad_cb_lag = acb_lag;
1841 p->subframe[index].ad_cb_gain = acb_gain;
1845 * Subtract the adaptive codebook contribution from the input
1846 * to obtain the residual.
1848 * @param buf target vector
1850 static void sub_acb_contrib(int16_t *residual, int16_t *impulse_resp,
1854 /* Subtract adaptive CB contribution to obtain the residual */
1855 for (i = 0; i < SUBFRAME_LEN; i++) {
1856 int64_t temp = buf[i] << 14;
1857 for (j = 0; j <= i; j++)
1858 temp -= residual[j] * impulse_resp[i - j];
1860 buf[i] = av_clipl_int32((temp << 2) + (1 << 15)) >> 16;
1865 * Quantize the residual signal using the fixed codebook (MP-MLQ).
1867 * @param optim optimized fixed codebook parameters
1868 * @param buf excitation vector
1870 static void get_fcb_param(FCBParam *optim, int16_t *impulse_resp,
1871 int16_t *buf, int pulse_cnt, int pitch_lag)
1874 int16_t impulse_r[SUBFRAME_LEN];
1875 int16_t temp_corr[SUBFRAME_LEN];
1876 int16_t impulse_corr[SUBFRAME_LEN];
1878 int ccr1[SUBFRAME_LEN];
1879 int ccr2[SUBFRAME_LEN];
1880 int amp, err, max, max_amp_index, min, scale, i, j, k, l;
1884 /* Update impulse response */
1885 memcpy(impulse_r, impulse_resp, sizeof(int16_t) * SUBFRAME_LEN);
1886 param.dirac_train = 0;
1887 if (pitch_lag < SUBFRAME_LEN - 2) {
1888 param.dirac_train = 1;
1889 gen_dirac_train(impulse_r, pitch_lag);
1892 for (i = 0; i < SUBFRAME_LEN; i++)
1893 temp_corr[i] = impulse_r[i] >> 1;
1895 /* Compute impulse response autocorrelation */
1896 temp = dot_product(temp_corr, temp_corr, SUBFRAME_LEN, 1);
1898 scale = normalize_bits_int32(temp);
1899 impulse_corr[0] = av_clipl_int32((temp << scale) + (1 << 15)) >> 16;
1901 for (i = 1; i < SUBFRAME_LEN; i++) {
1902 temp = dot_product(temp_corr + i, temp_corr, SUBFRAME_LEN - i, 1);
1903 impulse_corr[i] = av_clipl_int32((temp << scale) + (1 << 15)) >> 16;
1906 /* Compute crosscorrelation of impulse response with residual signal */
1908 for (i = 0; i < SUBFRAME_LEN; i++){
1909 temp = dot_product(buf + i, impulse_r, SUBFRAME_LEN - i, 1);
1911 ccr1[i] = temp >> -scale;
1913 ccr1[i] = av_clipl_int32(temp << scale);
1917 for (i = 0; i < GRID_SIZE; i++) {
1918 /* Maximize the crosscorrelation */
1920 for (j = i; j < SUBFRAME_LEN; j += GRID_SIZE) {
1921 temp = FFABS(ccr1[j]);
1924 param.pulse_pos[0] = j;
1928 /* Quantize the gain (max crosscorrelation/impulse_corr[0]) */
1931 max_amp_index = GAIN_LEVELS - 2;
1932 for (j = max_amp_index; j >= 2; j--) {
1933 temp = av_clipl_int32((int64_t)fixed_cb_gain[j] *
1934 impulse_corr[0] << 1);
1935 temp = FFABS(temp - amp);
1943 /* Select additional gain values */
1944 for (j = 1; j < 5; j++) {
1945 for (k = i; k < SUBFRAME_LEN; k += GRID_SIZE) {
1949 param.amp_index = max_amp_index + j - 2;
1950 amp = fixed_cb_gain[param.amp_index];
1952 param.pulse_sign[0] = (ccr2[param.pulse_pos[0]] < 0) ? -amp : amp;
1953 temp_corr[param.pulse_pos[0]] = 1;
1955 for (k = 1; k < pulse_cnt; k++) {
1957 for (l = i; l < SUBFRAME_LEN; l += GRID_SIZE) {
1960 temp = impulse_corr[FFABS(l - param.pulse_pos[k - 1])];
1961 temp = av_clipl_int32((int64_t)temp *
1962 param.pulse_sign[k - 1] << 1);
1964 temp = FFABS(ccr2[l]);
1967 param.pulse_pos[k] = l;
1971 param.pulse_sign[k] = (ccr2[param.pulse_pos[k]] < 0) ?
1973 temp_corr[param.pulse_pos[k]] = 1;
1976 /* Create the error vector */
1977 memset(temp_corr, 0, sizeof(int16_t) * SUBFRAME_LEN);
1979 for (k = 0; k < pulse_cnt; k++)
1980 temp_corr[param.pulse_pos[k]] = param.pulse_sign[k];
1982 for (k = SUBFRAME_LEN - 1; k >= 0; k--) {
1984 for (l = 0; l <= k; l++) {
1985 int prod = av_clipl_int32((int64_t)temp_corr[l] *
1986 impulse_r[k - l] << 1);
1987 temp = av_clipl_int32(temp + prod);
1989 temp_corr[k] = temp << 2 >> 16;
1992 /* Compute square of error */
1994 for (k = 0; k < SUBFRAME_LEN; k++) {
1996 prod = av_clipl_int32((int64_t)buf[k] * temp_corr[k] << 1);
1997 err = av_clipl_int32(err - prod);
1998 prod = av_clipl_int32((int64_t)temp_corr[k] * temp_corr[k]);
1999 err = av_clipl_int32(err + prod);
2003 if (err < optim->min_err) {
2004 optim->min_err = err;
2005 optim->grid_index = i;
2006 optim->amp_index = param.amp_index;
2007 optim->dirac_train = param.dirac_train;
2009 for (k = 0; k < pulse_cnt; k++) {
2010 optim->pulse_sign[k] = param.pulse_sign[k];
2011 optim->pulse_pos[k] = param.pulse_pos[k];
2019 * Encode the pulse position and gain of the current subframe.
2021 * @param optim optimized fixed CB parameters
2022 * @param buf excitation vector
2024 static void pack_fcb_param(G723_1_Subframe *subfrm, FCBParam *optim,
2025 int16_t *buf, int pulse_cnt)
2029 j = PULSE_MAX - pulse_cnt;
2031 subfrm->pulse_sign = 0;
2032 subfrm->pulse_pos = 0;
2034 for (i = 0; i < SUBFRAME_LEN >> 1; i++) {
2035 int val = buf[optim->grid_index + (i << 1)];
2037 subfrm->pulse_pos += combinatorial_table[j][i];
2039 subfrm->pulse_sign <<= 1;
2040 if (val < 0) subfrm->pulse_sign++;
2043 if (j == PULSE_MAX) break;
2046 subfrm->amp_index = optim->amp_index;
2047 subfrm->grid_index = optim->grid_index;
2048 subfrm->dirac_train = optim->dirac_train;
2052 * Compute the fixed codebook excitation.
2054 * @param buf target vector
2055 * @param impulse_resp impulse response of the combined filter
2057 static void fcb_search(G723_1_Context *p, int16_t *impulse_resp,
2058 int16_t *buf, int index)
2061 int pulse_cnt = pulses[index];
2064 optim.min_err = 1 << 30;
2065 get_fcb_param(&optim, impulse_resp, buf, pulse_cnt, SUBFRAME_LEN);
2067 if (p->pitch_lag[index >> 1] < SUBFRAME_LEN - 2) {
2068 get_fcb_param(&optim, impulse_resp, buf, pulse_cnt,
2069 p->pitch_lag[index >> 1]);
2072 /* Reconstruct the excitation */
2073 memset(buf, 0, sizeof(int16_t) * SUBFRAME_LEN);
2074 for (i = 0; i < pulse_cnt; i++)
2075 buf[optim.pulse_pos[i]] = optim.pulse_sign[i];
2077 pack_fcb_param(&p->subframe[index], &optim, buf, pulse_cnt);
2079 if (optim.dirac_train)
2080 gen_dirac_train(buf, p->pitch_lag[index >> 1]);
2084 * Pack the frame parameters into output bitstream.
2086 * @param frame output buffer
2087 * @param size size of the buffer
2089 static int pack_bitstream(G723_1_Context *p, unsigned char *frame, int size)
2092 int info_bits, i, temp;
2094 init_put_bits(&pb, frame, size);
2096 if (p->cur_rate == RATE_6300) {
2098 put_bits(&pb, 2, info_bits);
2101 put_bits(&pb, 8, p->lsp_index[2]);
2102 put_bits(&pb, 8, p->lsp_index[1]);
2103 put_bits(&pb, 8, p->lsp_index[0]);
2105 put_bits(&pb, 7, p->pitch_lag[0] - PITCH_MIN);
2106 put_bits(&pb, 2, p->subframe[1].ad_cb_lag);
2107 put_bits(&pb, 7, p->pitch_lag[1] - PITCH_MIN);
2108 put_bits(&pb, 2, p->subframe[3].ad_cb_lag);
2110 /* Write 12 bit combined gain */
2111 for (i = 0; i < SUBFRAMES; i++) {
2112 temp = p->subframe[i].ad_cb_gain * GAIN_LEVELS +
2113 p->subframe[i].amp_index;
2114 if (p->cur_rate == RATE_6300)
2115 temp += p->subframe[i].dirac_train << 11;
2116 put_bits(&pb, 12, temp);
2119 put_bits(&pb, 1, p->subframe[0].grid_index);
2120 put_bits(&pb, 1, p->subframe[1].grid_index);
2121 put_bits(&pb, 1, p->subframe[2].grid_index);
2122 put_bits(&pb, 1, p->subframe[3].grid_index);
2124 if (p->cur_rate == RATE_6300) {
2125 skip_put_bits(&pb, 1); /* reserved bit */
2127 /* Write 13 bit combined position index */
2128 temp = (p->subframe[0].pulse_pos >> 16) * 810 +
2129 (p->subframe[1].pulse_pos >> 14) * 90 +
2130 (p->subframe[2].pulse_pos >> 16) * 9 +
2131 (p->subframe[3].pulse_pos >> 14);
2132 put_bits(&pb, 13, temp);
2134 put_bits(&pb, 16, p->subframe[0].pulse_pos & 0xffff);
2135 put_bits(&pb, 14, p->subframe[1].pulse_pos & 0x3fff);
2136 put_bits(&pb, 16, p->subframe[2].pulse_pos & 0xffff);
2137 put_bits(&pb, 14, p->subframe[3].pulse_pos & 0x3fff);
2139 put_bits(&pb, 6, p->subframe[0].pulse_sign);
2140 put_bits(&pb, 5, p->subframe[1].pulse_sign);
2141 put_bits(&pb, 6, p->subframe[2].pulse_sign);
2142 put_bits(&pb, 5, p->subframe[3].pulse_sign);
2145 flush_put_bits(&pb);
2146 return frame_size[info_bits];
2149 static int g723_1_encode_frame(AVCodecContext *avctx, AVPacket *avpkt,
2150 const AVFrame *frame, int *got_packet_ptr)
2152 G723_1_Context *p = avctx->priv_data;
2153 int16_t unq_lpc[LPC_ORDER * SUBFRAMES];
2154 int16_t qnt_lpc[LPC_ORDER * SUBFRAMES];
2155 int16_t cur_lsp[LPC_ORDER];
2156 int16_t weighted_lpc[LPC_ORDER * SUBFRAMES << 1];
2157 int16_t vector[FRAME_LEN + PITCH_MAX];
2159 int16_t *in = (const int16_t *)frame->data[0];
2164 highpass_filter(in, &p->hpf_fir_mem, &p->hpf_iir_mem);
2166 memcpy(vector, p->prev_data, HALF_FRAME_LEN * sizeof(int16_t));
2167 memcpy(vector + HALF_FRAME_LEN, in, FRAME_LEN * sizeof(int16_t));
2169 comp_lpc_coeff(vector, unq_lpc);
2170 lpc2lsp(&unq_lpc[LPC_ORDER * 3], p->prev_lsp, cur_lsp);
2171 lsp_quantize(p->lsp_index, cur_lsp, p->prev_lsp);
2174 memcpy(vector + LPC_ORDER, p->prev_data + SUBFRAME_LEN,
2175 sizeof(int16_t) * SUBFRAME_LEN);
2176 memcpy(vector + LPC_ORDER + SUBFRAME_LEN, in,
2177 sizeof(int16_t) * (HALF_FRAME_LEN + SUBFRAME_LEN));
2178 memcpy(p->prev_data, in + HALF_FRAME_LEN,
2179 sizeof(int16_t) * HALF_FRAME_LEN);
2180 memcpy(in, vector + LPC_ORDER, sizeof(int16_t) * FRAME_LEN);
2182 perceptual_filter(p, weighted_lpc, unq_lpc, vector);
2184 memcpy(in, vector + LPC_ORDER, sizeof(int16_t) * FRAME_LEN);
2185 memcpy(vector, p->prev_weight_sig, sizeof(int16_t) * PITCH_MAX);
2186 memcpy(vector + PITCH_MAX, in, sizeof(int16_t) * FRAME_LEN);
2188 scale_vector(vector, FRAME_LEN + PITCH_MAX);
2190 p->pitch_lag[0] = estimate_pitch(vector, PITCH_MAX);
2191 p->pitch_lag[1] = estimate_pitch(vector, PITCH_MAX + HALF_FRAME_LEN);
2193 for (i = PITCH_MAX, j = 0; j < SUBFRAMES; i += SUBFRAME_LEN, j++)
2194 comp_harmonic_coeff(vector + i, p->pitch_lag[j >> 1], hf + j);
2196 memcpy(vector, p->prev_weight_sig, sizeof(int16_t) * PITCH_MAX);
2197 memcpy(vector + PITCH_MAX, in, sizeof(int16_t) * FRAME_LEN);
2198 memcpy(p->prev_weight_sig, vector + FRAME_LEN, sizeof(int16_t) * PITCH_MAX);
2200 for (i = 0, j = 0; j < SUBFRAMES; i += SUBFRAME_LEN, j++)
2201 harmonic_filter(hf + j, vector + PITCH_MAX + i, in + i);
2203 inverse_quant(cur_lsp, p->prev_lsp, p->lsp_index, 0);
2204 lsp_interpolate(qnt_lpc, cur_lsp, p->prev_lsp);
2206 memcpy(p->prev_lsp, cur_lsp, sizeof(int16_t) * LPC_ORDER);
2209 for (i = 0; i < SUBFRAMES; i++) {
2210 int16_t impulse_resp[SUBFRAME_LEN];
2211 int16_t residual[SUBFRAME_LEN + PITCH_ORDER - 1];
2212 int16_t flt_in[SUBFRAME_LEN];
2213 int16_t zero[LPC_ORDER], fir[LPC_ORDER], iir[LPC_ORDER];
2216 * Compute the combined impulse response of the synthesis filter,
2217 * formant perceptual weighting filter and harmonic noise shaping filter
2219 memset(zero, 0, sizeof(int16_t) * LPC_ORDER);
2220 memset(vector, 0, sizeof(int16_t) * PITCH_MAX);
2221 memset(flt_in, 0, sizeof(int16_t) * SUBFRAME_LEN);
2223 flt_in[0] = 1 << 13; /* Unit impulse */
2224 synth_percept_filter(qnt_lpc + offset, weighted_lpc + (offset << 1),
2225 zero, zero, flt_in, vector + PITCH_MAX, 1);
2226 harmonic_filter(hf + i, vector + PITCH_MAX, impulse_resp);
2228 /* Compute the combined zero input response */
2230 memcpy(fir, p->perf_fir_mem, sizeof(int16_t) * LPC_ORDER);
2231 memcpy(iir, p->perf_iir_mem, sizeof(int16_t) * LPC_ORDER);
2233 synth_percept_filter(qnt_lpc + offset, weighted_lpc + (offset << 1),
2234 fir, iir, flt_in, vector + PITCH_MAX, 0);
2235 memcpy(vector, p->harmonic_mem, sizeof(int16_t) * PITCH_MAX);
2236 harmonic_noise_sub(hf + i, vector + PITCH_MAX, in);
2238 acb_search(p, residual, impulse_resp, in, i);
2239 gen_acb_excitation(residual, p->prev_excitation,p->pitch_lag[i >> 1],
2240 p->subframe[i], p->cur_rate);
2241 sub_acb_contrib(residual, impulse_resp, in);
2243 fcb_search(p, impulse_resp, in, i);
2245 /* Reconstruct the excitation */
2246 gen_acb_excitation(impulse_resp, p->prev_excitation, p->pitch_lag[i >> 1],
2247 p->subframe[i], RATE_6300);
2249 memmove(p->prev_excitation, p->prev_excitation + SUBFRAME_LEN,
2250 sizeof(int16_t) * (PITCH_MAX - SUBFRAME_LEN));
2251 for (j = 0; j < SUBFRAME_LEN; j++)
2252 in[j] = av_clip_int16((in[j] << 1) + impulse_resp[j]);
2253 memcpy(p->prev_excitation + PITCH_MAX - SUBFRAME_LEN, in,
2254 sizeof(int16_t) * SUBFRAME_LEN);
2256 /* Update filter memories */
2257 synth_percept_filter(qnt_lpc + offset, weighted_lpc + (offset << 1),
2258 p->perf_fir_mem, p->perf_iir_mem,
2259 in, vector + PITCH_MAX, 0);
2260 memmove(p->harmonic_mem, p->harmonic_mem + SUBFRAME_LEN,
2261 sizeof(int16_t) * (PITCH_MAX - SUBFRAME_LEN));
2262 memcpy(p->harmonic_mem + PITCH_MAX - SUBFRAME_LEN, vector + PITCH_MAX,
2263 sizeof(int16_t) * SUBFRAME_LEN);
2266 offset += LPC_ORDER;
2269 if ((ret = ff_alloc_packet2(avctx, avpkt, 24)))
2272 *got_packet_ptr = 1;
2273 avpkt->size = pack_bitstream(p, avpkt->data, avpkt->size);
2277 AVCodec ff_g723_1_encoder = {
2279 .type = AVMEDIA_TYPE_AUDIO,
2280 .id = AV_CODEC_ID_G723_1,
2281 .priv_data_size = sizeof(G723_1_Context),
2282 .init = g723_1_encode_init,
2283 .encode2 = g723_1_encode_frame,
2284 .long_name = NULL_IF_CONFIG_SMALL("G.723.1"),
2285 .sample_fmts = (const enum AVSampleFormat[]){AV_SAMPLE_FMT_S16,
2286 AV_SAMPLE_FMT_NONE},